Abstract
We investigate the problem of gambling with uncertainty in outcome probabilities. Stochastic optimization models are proposed for optimal investing on events with mutually exclusive outcomes when probabilities are estimated using multinomial logistic regression. Special attention is given to the case of there being two outcomes, and the general case of many outcomes. An empirical study using simulated data was conducted where the loss of return from probability estimation error is observed, and superior returns are achieved taking it into consideration.
Introduction
The classic Kelly Criterion is widely regarded as the gold standard for long-term wealth growth in gambling and investment. By calculating the optimal fraction of a bankroll to wager based on known odds and probabilities, it maximizes the expected logarithm of wealth. However, the formula possesses a significant practical flaw: it assumes that the probability estimates provided to it are perfect.
In real-world scenarios, such as horse racing, probabilities are rarely known with certainty; they are estimated using statistical models like multinomial logistic regression. Because the Kelly formula is inherently aggressive, it is highly sensitive to errors in these estimates. If a model overestimates an outcome's probability, the Kelly Criterion will recommend an overbet. Over time, these errors lead to increased volatility, lower-than-expected returns, and in extreme cases, total ruin. While practitioners often use "Fractional Kelly" (betting a fixed percentage of the recommended amount) as a safety margin, this paper proposes a more mathematically rigorous framework to account for estimation noise.
Methodology
Modeling Uncertainty
Rather than relying on point estimates, this research treats winning probabilities as random variables. The estimation process utilizes Multinomial Logistic Regression to derive win probabilities based on various horse-specific factors. To account for potential model misspecification—where the model might fail to capture every nuance of a real race—the author employs a Sandwich Covariance Matrix. This provides a robust measure of parameter uncertainty, acknowledging that the model is merely an approximation of reality.
Proposed Optimization Models
The paper introduces several stochastic optimization models designed to handle this uncertainty:
- Expected Monte Carlo (Emc): This model maximizes the expected log wealth across the entire distribution of possible probabilities, rather than just the average probability.
- Lower Bound (Elb): Utilizing Jensen’s inequality, this approach establishes a conservative "floor" for probabilities to prevent aggressive overbetting.
- Chance-Constrained (CC): A "safety-first" strategy that ensures the bet achieves a specific return threshold in at least $(1-\alpha)\%$ of scenarios.
- Expected Chance-Constrained (ECC): A hybrid approach that seeks to maximize expected returns while maintaining a chance constraint to prevent negative outcomes in the face of high uncertainty.
Experimental Setup
To validate these models, the study utilizes simulated data rather than historical racing data. The primary advantage of simulation is that the "true" underlying probability of a horse winning is known to the researcher. This allows for a direct comparison between the models' performance and the theoretical maximum return (the "True Kelly"). The author conducted 2,500 trials across four race configurations, ranging from 2 to 30 horses, to test the robustness of each strategy.
Key Insights and Results
The Overestimation Trap
A critical finding of the study is the "Overestimation Trap." In a noisy environment, the bets that appear most attractive to a standard Kelly model are often those where the estimation error is most positive. Consequently, the standard Kelly model naturally gravitates toward the most "wrong" predictions, leading to systematic overbetting.
Performance Comparison
The empirical study compared the proposed models against Standard Kelly (S), Fractional Kelly (F), and True Kelly (T).
- Superiority of ECC: The Expected Chance-Constrained (ECC) model with $\alpha = 0.40$ emerged as the top performer. It consistently outperformed the standard Kelly model across all experiments.
- Total Returns: In the aggregate results, the Standard Kelly model achieved a return of 18.134, while the ECC model reached 18.484.
- Inefficiency of Fractional Kelly: While the 50% Fractional Kelly strategy is safer, its total return was significantly lower (13.268). This suggests that mathematically modeling uncertainty is far more efficient than applying a blunt reduction to bet sizes.
- The Cost of Ignorance: Despite the improvements, a large gap remains between the ECC model (18.484) and the True Kelly return (27.463). This gap represents the inherent "alpha" lost due to the inability to perfectly predict future outcomes.
Conclusion
By incorporating uncertainty directly into the optimization process, investors can mitigate the risks associated with model error. The parameter $\alpha$ in chance-constrained models serves as a vital tuning knob for risk; while a low $\alpha$ provides safety, being overly conservative can stifle growth as much as being too aggressive.
Future research may extend these methodologies to financial markets, particularly for assets following Geometric Brownian Motion, where parameter uncertainty regarding drift and volatility remains a primary challenge for portfolio managers.
Summary updated